A shield tunneling stratum intelligent identification method and system based on a strict single-sample prototype network and a verification set scale calibration

Abstract

The application discloses a kind of based on strict single sample prototype network and verification set scale calibration shield tunneling stratum intelligent identification method and system, first, according to the division of training set, verification set and test set from the sample that is collected in real time from shield tunneling machine according to predetermined time and engineering interval, then the training set is used to normalize the preprocessing of multidimensional mechanism parameter, then the cosine interval classifier is used to supervise the pre-training of lightweight MLP encoder and obtain discriminative embedding space;The application realizes the online identification function of high precision, reproducible and low latency to shield tunneling stratum under the condition of few labels, not only can consider overall accuracy and minority class recall under the premise of not increasing labeling cost, so that the output probability is more calibrated, but also has scalable and fast deployment capability, significantly improves the real-time and reliability of shield tunneling stratum identification.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) stratum identification technology, specifically to an intelligent TBM stratum identification method and system based on rigorous single-sample prototype network and validation set scale calibration. Background Technology

[0002] The geological type of the tunnel boring machine directly affects the dynamic distribution of propulsion parameters, which include cutterhead rotation speed / torque, propulsion speed, penetration depth, total thrust, chamber pressure, screw conveyor torque, and thrust factor.

[0003] Currently, most existing shield tunneling projects use on-site annotation to identify the strata in the tunnel. This results in traditional supervised learning having poor performance in identifying the strata categories in shield tunneling. Furthermore, the cost of strata annotation is high, and the identification of strata categories is uneven. Although convolutional or sequence models can fit the data, they are not adaptable to small sample labels and incremental labels. This not only makes the identification of strata in shield tunneling more costly, but also leads to unstable support sample selection and leakage of training and validation information. Therefore, it is necessary to design an intelligent identification method and system for shield tunneling strata based on a strictly single-sample prototype network and validation set scale calibration. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies. To better and more effectively address the problems in current shield tunneling projects, where most rely on on-site annotation for identifying the strata, traditional supervised learning methods are ineffective in identifying strata categories. Furthermore, the high cost of strata annotation and the uneven distribution of strata categories, coupled with the fact that convolutional or sequence models, while capable of fitting the data, lack adaptability to small sample labels and incremental labels, not only increase the cost of strata identification but also lead to unstable support sample selection and leakage of training and validation information. This invention provides an intelligent strata identification method and system for shield tunneling based on a rigorous single-sample prototype network and validation set scale calibration. It achieves high-precision, reproducible, and low-latency online identification of strata under minimal label conditions. It not only balances overall accuracy and minority class recall for more calibrated output probabilities without increasing annotation costs but also possesses scalability and rapid deployment capabilities, significantly improving the real-time performance and reliability of strata identification for shield tunneling.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for intelligent identification of geological formations during tunnel boring machines (TBM) construction, based on rigorous single-sample prototype networks and validation set scale calibration, includes the following steps:

[0007] Step A involves dividing the samples collected in real time from the tunnel boring machine into training set, validation set and test set according to a predetermined time and engineering interval, and then using the training set to perform normalization preprocessing on the multidimensional mechanism parameters.

[0008] Step B involves using a cosine interval classifier to perform supervised pre-training on the lightweight MLP encoder and obtain the discriminative embedding space.

[0009] Step C: Select a unique supporting sample for each class from the training set within the embedding space according to the class center proximal sample strategy;

[0010] Step D involves keeping the support samples unchanged in each training task and sampling query samples from the training set in a class-balanced manner. Then, an upper limit is set on the number of times a single sample is used in a single training period, and temperature-scaled cosine similarity cross-entropy is used to optimize the metric space.

[0011] Step E: On the validation set, scale calibration is performed using the first grid search scale to obtain the second grid search scale, and then strict single-sample classification is performed on the test set using the second grid search scale to obtain the classification result.

[0012] Step F involves processing the classification results using the softmax function to obtain the probability of the identified category, then averaging the probabilities using the identified category dimension and applying... The function obtains the final recognition result.

[0013] The aforementioned intelligent identification method for tunnel boring machine (TBM) strata based on rigorous single-sample prototype networks and validation set scale calibration involves step A, which divides samples collected in real-time from the TBM into training, validation, and test sets according to predetermined time and engineering intervals. The training set is then used to perform normalized preprocessing on the multi-dimensional mechanistic parameters. The specific steps are as follows.

[0014] Step A1 involves dividing the samples collected in real time from the tunnel boring machine into a training set, a validation set, and a test set based on a predetermined time and project interval. The training set is used to fit the Min–Max normalized parameters, as shown in formula (1).

[0015] (1)

[0016] in, The original value of the sample on the d-th feature. The normalized value of the sample on the d-th feature. Let be the minimum value of the d-th feature in the training set. Let be the maximum value of the d-th feature in the training set;

[0017] Step A2 involves using the training set to perform normalization preprocessing on the multidimensional mechanism parameters, which include cutterhead speed, cutterhead torque, propulsion speed, penetration depth, total thrust, earth pressure, screw conveyor torque, and thrust factor. The normalization preprocessing includes label encoding, anomaly filling, and leakage prevention normalization.

[0018] The aforementioned intelligent geological identification method for tunnel boring machines based on a strictly single-sample prototype network and validation set scale calibration, step B, involves using a cosine interval classifier to perform supervised pre-training of a lightweight MLP encoder to obtain a discriminative embedding space. The specific steps are as follows.

[0019] Step B1: Construct a lightweight MLP encoder and assign a cosine classification head, as shown in formula (2).

[0020] ;

[0021] ;

[0022] (2)

[0023] in, The L2-normalized sample embedding vector For MLP encoders, The input feature vector is the normalized form. As a learning standard, Cosine intervals, The class weight vector after L2 normalization. This is the original class weight vector;

[0024] Step B2, calculate the weighted cross-entropy loss. As shown in formula (3),

[0025] (3)

[0026] in, For sample index, and All are category indexes. Let n be the true class label of the nth sample. and These are the numerator and denominator of the softmax function, respectively.

[0027] The aforementioned intelligent geological formation identification method for tunnel boring machines based on a strictly single-sample prototype network and validation set scale calibration, in step C, involves selecting a unique supporting sample for each class from the training set within the embedding space according to a class-center-proximal sample strategy. The specific steps are as follows.

[0028] Step C1: Embed and normalize the training set samples, then calculate the class mean for each class. As shown in formula (4),

[0029] ;

[0030] (4)

[0031] in, Let be the normalized embedding vector of the nth training sample. This represents the selection of support samples for class c from the training set. For transpose;

[0032] Step C2 involves selecting a unique supporting sample for each class, as shown in formula (5).

[0033] (5)

[0034] in, For class c, the fixed support prototype vector. Selected as supporting samples The normalized features of the training samples.

[0035] The aforementioned intelligent geological formation identification method for tunnel boring machines based on a strictly single-sample prototype network and validation set scale calibration, in step D, involves keeping the support samples unchanged in each training task and sampling query samples from the training set in a class-balanced manner. Then, an upper limit is set on the number of times a single sample is used for query samples within a single training period, and the metric space is optimized using temperature-scaled cosine similarity cross-entropy. The specific steps are as follows.

[0036] Step D1: In each training task, the support samples are kept unchanged and the query samples are sampled uniformly from the training set by class. During the sampling process, uniform sampling without replacement is used first, and padding with replacement is performed when there is insufficient sample.

[0037] Step D2 sets an upper limit on the number of times a single sample can be used in a single training period, as shown in formula (6).

[0038] (6)

[0039] in, To query the number of times the sample was used, The maximum number of times each query sample is used within a single training task;

[0040] Step D3 involves optimizing the metric space using temperature-scaled cosine similarity cross-entropy, as shown in Equation (7).

[0041] ;

[0042] (7)

[0043] in, Let be the weighted cosine similarity of the i-th query sample with respect to category c. The first grid search scale, Let be the normalized embedding vector of the i-th query sample. The loss function for each training task, For query set, To query the sample index, For the i-th query sample in the true category The indexation results on, This represents the indexed result of the i-th query sample in the true class c.

[0044] The aforementioned intelligent identification method for tunnel boring machine (TBM) strata based on strict single-sample prototype network and validation set scale calibration, step E, involves performing scale calibration on the validation set using a first grid search scale to obtain a second grid search scale, and then performing strict single-sample classification on the test set using the second grid search scale to obtain the classification result. The specific steps are as follows.

[0045] Step E1: Scale calibration is performed on the validation set using the first grid search scale to obtain the second grid search scale, as shown in formula (8).

[0046] (8)

[0047] in, This is the second grid search scale. For the candidate scale set, To improve classification accuracy, To verify the embedding vector, The category prototype matrix;

[0048] Step E2: Perform strict single-sample classification on the test set using the second grid search scale and obtain the classification results, as shown in formula (9).

[0049] (9)

[0050] in, For the class label of the input sample x, Let x be the normalized embedding vector of sample x.

[0051] The aforementioned intelligent geological identification method for tunnel boring machines based on a strictly single-sample prototype network and validation set scale calibration, in step F, involves processing the classification results using a softmax function to obtain the probability of the identified category, then averaging the probabilities using the dimensions of the identified categories and applying... The function obtains the final recognition result, as shown in formula (10).

[0052] (10)

[0053] in, For the final identification result, To identify the probability of the category.

[0054] A shield tunneling stratum intelligent identification system based on strict single-sample prototype network and validation set scale calibration includes a preprocessing module, a cosine interval classification module, a fixed selection module, an equal sampling module, a strict single-sample classification module, and a probability averaging module. The preprocessing module is used to divide samples collected in real time from the shield machine into training set, validation set, and test set according to a predetermined time and engineering interval, and then use the training set to perform normalized preprocessing on multidimensional mechanism parameters.

[0055] The cosine interval classification module is used to perform supervised pre-training of the lightweight MLP encoder using a cosine interval classifier and obtain a discriminative embedding space.

[0056] The fixed selection module is used to select a unique supporting sample for each class from the training set within the embedding space according to the class center proximal sample strategy;

[0057] The equalization sampling module is used to keep the support samples unchanged in each training task and to equally sample query samples from the training set by class. It also sets an upper limit on the number of times a single sample is used in a single training period and uses temperature-scaled cosine similarity cross-entropy to optimize the metric space.

[0058] The strict single-sample classification module is used to perform scale calibration on the validation set using a first grid search scale and obtain a second grid search scale, and then perform strict single-sample classification on the test set using the second grid search scale and obtain the classification result.

[0059] The probability averaging module is used to process the classification results using the softmax function to obtain the probability of the identified category, and then to perform an arithmetic average using the identified category dimension. The function obtains the final recognition result.

[0060] The beneficial effects of this invention are as follows: This invention provides a method and system for intelligent identification of geological formations during tunnel boring machine (TBM) excavation based on a rigorous single-sample prototype network and validation set scale calibration. First, samples collected in real-time from the TBM are divided into training, validation, and test sets according to predetermined time and engineering intervals. Then, the training set is used to standardize and preprocess multi-dimensional mechanism parameters. Next, a cosine interval classifier is used to supervise the pre-training of a lightweight MLP encoder to obtain a discriminative embedding space. Within the embedding space, a unique support sample is selected from the training set for each class based on a class-centered near-end sample strategy. Subsequently, in each training task, the support samples are kept constant, and query samples are sampled from the training set in a class-balanced manner. An upper limit is set on the number of times a single sample is used for querying within a single training period, and temperature-scaled cosine similarity cross-entropy is used to optimize the metric space. Then, a first grid search scale is used on the validation set to perform scale calibration and obtain a second grid search scale. Finally, a rigorous single-sample classification is performed on the test set using the second grid search scale to obtain the classification results. Finally, the classification results are processed using a softmax function to obtain the probability of the identified class. An arithmetic mean is then calculated using the identification class dimension, and... The function obtains the final recognition result; this invention effectively realizes the intelligent identification method and system for shield tunneling strata, which has the function of high-precision, reproducible and low-latency online identification of shield tunneling strata under conditions with very few labels. Moreover, for multi-dimensional mechanism parameters such as cutterhead rotation speed, cutterhead torque, propulsion speed, penetration depth, total thrust, earth pressure, screw conveyor torque and thrust factor, leakage prevention normalization, robust missing and anomaly filling can be performed by a training set fitting method. Furthermore, by constructing a lightweight MLP encoder and performing supervised pre-training with a cosine interval classification head to shape the discriminative embedding, the embedding space is classified according to the proximal end of the class center. With a fixed and unique sample, strict one-shot fine-tuning is carried out by combining class-balanced set-by-set sampling and query quota constraints, and the metric space is optimized with temperature-scaled cosine similarity. At the same time, in the validation phase, only the logit scale factor is searched on the validation set to complete the temperature and scale calibration. In the testing phase, one training support sample per class plus the optimal scale is used to identify the strata in shield tunneling. This not only balances overall accuracy and minority class recall without increasing the annotation cost, making the output probability more calibrated, but also has scalability and rapid deployment capabilities, significantly improving the real-time performance and reliability of shield tunneling strata identification. Attached Figure Description

[0061] Figure 1 This is an overall flowchart of a shield tunneling stratum intelligent identification method based on a rigorous single-sample prototype network and validation set scale calibration according to the present invention. Detailed Implementation

[0062] The present invention will now be further described with reference to the accompanying drawings.

[0063] like Figure 1 As shown, the present invention provides a method for intelligent identification of geological formations during tunnel boring machine (TBM) excavation based on rigorous single-sample prototype network and validation set scale calibration, comprising the following steps:

[0064] Step A involves dividing the samples collected in real time from the tunnel boring machine into training, validation, and test sets based on predetermined time periods and engineering intervals. The training set is then used to perform normalization preprocessing on the multidimensional mechanism parameters. The specific steps are as follows.

[0065] Step A1 involves dividing the samples collected in real time from the tunnel boring machine into a training set, a validation set, and a test set based on a predetermined time and project interval. The training set is used to fit the Min–Max normalized parameters, as shown in formula (1).

[0066] (1)

[0067] in, The original value of the sample on the d-th feature. The normalized value of the sample on the d-th feature. Let be the minimum value of the d-th feature in the training set. Let be the maximum value of the d-th feature in the training set;

[0068] Step A2 involves using the training set to perform normalization preprocessing on the multidimensional mechanism parameters, which include cutterhead speed, cutterhead torque, propulsion speed, penetration depth, total thrust, earth pressure, screw conveyor torque, and thrust factor. The normalization preprocessing includes label encoding, anomaly filling, and leakage prevention normalization.

[0069] Step B involves using a cosine-space classifier to perform supervised pre-training on the lightweight MLP encoder and obtain the discriminative embedding space. The specific steps are as follows:

[0070] Step B1: Construct a lightweight MLP encoder and assign a cosine classification head, as shown in formula (2).

[0071] ;

[0072] ;

[0073] (2)

[0074] in, The L2-normalized sample embedding vector For MLP encoders, The input feature vector is the normalized form. As a learning standard, Cosine intervals, The class weight vector after L2 normalization. This is the original class weight vector;

[0075] Step B2, calculate the weighted cross-entropy loss. As shown in formula (3),

[0076] (3)

[0077] in, For sample index, and All are category indexes. Let n be the true class label of the nth sample. and These are the numerator and denominator of the softmax function, respectively.

[0078] Step C involves selecting a unique supporting sample for each class from the training set within the embedding space based on a class-center-proximal sample strategy. The specific steps are as follows.

[0079] Step C1: Embed and normalize the training set samples, then calculate the class mean for each class. As shown in formula (4),

[0080] ;

[0081] (4)

[0082] in, Let be the normalized embedding vector of the nth training sample. This represents the selection of support samples for class c from the training set. For transpose;

[0083] Step C2 involves selecting a unique supporting sample for each class, as shown in formula (5).

[0084] (5)

[0085] in, For class c, the fixed support prototype vector. Selected as supporting samples The normalized features of the training samples.

[0086] Step D involves keeping the support samples constant in each training task and sampling query samples from the training set in a class-balanced manner. An upper limit is set on the number of times a single sample is used in query samples within a single training period, and the metric space is optimized using temperature-scaled cosine similarity cross-entropy. The specific steps are as follows.

[0087] Step D1: In each training task, the support samples are kept unchanged and the query samples are sampled uniformly from the training set by class. During the sampling process, uniform sampling without replacement is used first, and padding with replacement is performed when there is insufficient sample.

[0088] Step D2 sets an upper limit on the number of times a single sample can be used in a single training period, as shown in formula (6).

[0089] (6)

[0090] in, To query the number of times the sample was used, The maximum number of times each query sample is used within a single training task;

[0091] Step D3 involves optimizing the metric space using temperature-scaled cosine similarity cross-entropy, as shown in Equation (7).

[0092] ;

[0093] (7)

[0094] in, Let be the weighted cosine similarity of the i-th query sample with respect to category c. The first grid search scale, Let be the normalized embedding vector of the i-th query sample. The loss function for each training task, For query set, To query the sample index, For the i-th query sample in the true category The indexation results on, This represents the indexed result of the i-th query sample in the true class c.

[0095] Step E involves performing scale calibration on the validation set using the first grid search scale to obtain the second grid search scale, and then performing rigorous single-sample classification on the test set using the second grid search scale to obtain the classification results. The specific steps are as follows:

[0096] Step E1: Scale calibration is performed on the validation set using the first grid search scale to obtain the second grid search scale, as shown in formula (8).

[0097] (8)

[0098] in, This is the second grid search scale. For the candidate scale set, To improve classification accuracy, To verify the embedding vector, The category prototype matrix;

[0099] Step E2: Perform strict single-sample classification on the test set using the second grid search scale and obtain the classification results, as shown in formula (9).

[0100] (9)

[0101] in, For the class label of the input sample x, Let x be the normalized embedding vector of sample x.

[0102] Step F involves processing the classification results using the softmax function to obtain the probability of the identified category, then averaging the probabilities using the identified category dimension and applying... The function obtains the final recognition result, as shown in formula (10).

[0103] (10)

[0104] in, For the final identification result, To identify the probability of the category.

[0105] A shield tunneling geological formation intelligent identification system based on rigorous single-sample prototype network and validation set scale calibration includes a preprocessing module, a cosine interval classification module, a fixed selection module, a balanced sampling module, a rigorous single-sample classification module, and a probability averaging module. The preprocessing module divides samples collected in real-time from the shield machine into training, validation, and test sets according to predetermined time and engineering intervals, and then uses the training set to perform normalized preprocessing on multidimensional mechanism parameters. The cosine interval classification module uses a cosine interval classifier to perform a lightweight MLP. The encoder undergoes supervised pre-training to obtain a discriminative embedding space. The fixed selection module selects a unique support sample for each class from the training set within the embedding space based on a class-centered near-end sample strategy. The balanced sampling module maintains the support samples unchanged in each training task and evenly samples query samples from the training set by class. It then sets an upper limit on the number of times a single sample is used for query samples within a single training period and optimizes the metric space using temperature-scaled cosine similarity cross-entropy. The strict single-sample classification module performs scale calibration on the validation set using a first grid search scale to obtain a second grid search scale, and then performs strict single-sample classification on the test set using the second grid search scale to obtain the classification result. The probability averaging module processes the classification result using a softmax function to obtain the probability of the identified class, then performs an arithmetic average using the identified class dimension and applies... The function obtains the final recognition result.

[0106] To better illustrate the effectiveness of this invention, a specific embodiment of the intelligent identification method and system for tunnel boring machine (TBM) strata in subway tunneling is described below.

[0107] This embodiment conducts experimental verification on urban subway tunnel boring machine (TBM) monitoring data. The subway TBM monitoring dataset contains 30,000 labeled samples, covering silty sand, mudstone, and gravel strata. While maintaining consistency in field measurements and filtering under conventional working conditions, derived and sliding window statistical features were constructed from the original quantities of propulsion speed, cutterhead rotation speed and torque, total propulsion force, earth pressure, and auger conveyor indicators, and then uniformly normalized. In this embodiment, under a strict single-sample setting, only a single sample is fixed as the prototype for each category, and inferences are then made solely based on this fixed prototype. As shown in Table 1, the intelligent stratum identification method and system for TBM tunneling of this invention can stably and accurately track stratum changes in the TBM tunneling time-series characteristics.

[0108] Table 1. Results of Time-Sequence Feature Identification for Shield Tunneling;

[0109]

[0110] As shown in the confusion matrix in Table 2, most mainstream stratigraphic categories achieved high accuracy, while exhibiting ideal recall for gravel strata, demonstrating identification ability and robustness under extremely low sample conditions.

[0111] Table 2. Results of confusion in the identification of tunnel boring machine time sequence features;

[0112]

[0113] In summary, the intelligent identification system for shield tunneling strata based on a rigorous single-sample prototype network and validation set scale calibration of the present invention achieves accurate identification and continuous tracking of shield tunneling strata without increasing annotation costs through an integrated technical approach that combines rich original, derived, and sliding window features, cosine interval pre-trained shaping embedding, fixed single-sample prototype, rigorous single-sample training fine-tuning, and validation set temperature and scale calibration. Compared with conventional classifiers that rely on a large number of samples, the present invention can significantly improve recall rate and overall stability in a few strata and has the engineering advantages of being lightweight, transferable, and interpretable, thereby meeting the online decision-making needs of shield tunneling monitoring data.

[0114] In summary, the present invention provides a method and system for intelligent identification of geological formations in tunnel boring machines (TBMs) based on a strictly single-sample prototype network and validation set scale calibration. First, samples collected in real-time from the TBM are divided into training, validation, and test sets according to predetermined time and engineering intervals. Then, the training set is used to standardize and preprocess multi-dimensional mechanistic parameters. Next, a cosine interval classifier is used to supervise the pre-training of a lightweight MLP encoder to obtain a discriminative embedding space. Within the embedding space, a unique support sample is selected from the training set for each class according to a class-centered near-end sample strategy. Subsequently, in each training task, the support samples are kept constant, and query samples are sampled from the training set in a class-balanced manner. An upper limit is set on the number of times a single sample is used for querying within a single training period, and temperature-scaled cosine similarity cross-entropy is used to optimize the metric space. Then, a first grid search scale is used for scale calibration on the validation set to obtain a second grid search scale. Finally, a strict single-sample classification is performed on the test set using the second grid search scale to obtain classification results. Finally, the classification results are processed using a softmax function to obtain the probability of the identified class, and an arithmetic mean is calculated using the identification class dimension. The function obtains the final recognition result; this invention effectively realizes the intelligent identification method and system for shield tunneling strata, which has the function of high-precision, reproducible and low-latency online identification of shield tunneling strata under conditions with very few labels. Moreover, for multi-dimensional mechanism parameters such as cutterhead rotation speed, cutterhead torque, propulsion speed, penetration depth, total thrust, earth pressure, screw conveyor torque and thrust factor, leakage prevention normalization, robust missing and anomaly filling can be performed by a training set fitting method. Furthermore, by constructing a lightweight MLP encoder and performing supervised pre-training with a cosine interval classification head to shape the discriminative embedding, the embedding space is classified according to the proximal end of the class center. With a fixed and unique sample, strict one-shot fine-tuning is carried out by combining class-balanced set-by-set sampling and query quota constraints, and the metric space is optimized with temperature-scaled cosine similarity. At the same time, in the validation phase, only the logit scale factor is searched on the validation set to complete the temperature and scale calibration. In the testing phase, one training support sample per class plus the optimal scale is used to identify the strata in shield tunneling. This not only balances overall accuracy and minority class recall without increasing the annotation cost, making the output probability more calibrated, but also has scalability and rapid deployment capabilities, significantly improving the real-time performance and reliability of shield tunneling strata identification.

[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of geological formations during tunnel boring machines (TBM) operation based on rigorous single-sample prototype networks and validation set scale calibration, characterized in that: Includes the following steps, Step A involves dividing the samples collected in real time from the tunnel boring machine into training set, validation set and test set according to a predetermined time and engineering interval, and then using the training set to perform normalization preprocessing on the multidimensional mechanism parameters. Step B involves using a cosine interval classifier to perform supervised pre-training on the lightweight MLP encoder and obtain the discriminative embedding space. Step C: Select a unique supporting sample for each class from the training set within the embedding space according to the class center proximal sample strategy; Step D involves keeping the support samples unchanged in each training task and sampling query samples from the training set in a class-balanced manner. Then, an upper limit is set on the number of times a single sample is used in a single training period, and temperature-scaled cosine similarity cross-entropy is used to optimize the metric space. Step E: On the validation set, scale calibration is performed using the first grid search scale to obtain the second grid search scale, and then strict single-sample classification is performed on the test set using the second grid search scale to obtain the classification result. Step F involves processing the classification results using the softmax function to obtain the probability of the identified category, then averaging the probabilities using the identified category dimension and applying... The function obtains the final recognition result.

2. The intelligent identification method for tunnel boring machine strata based on strict single-sample prototype network and validation set scale calibration as described in claim 1, characterized in that: Step A involves dividing the samples collected in real time from the tunnel boring machine into training, validation, and test sets based on predetermined time periods and engineering intervals. The training set is then used to perform normalization preprocessing on the multidimensional mechanism parameters. The specific steps are as follows. Step A1 involves dividing the samples collected in real time from the tunnel boring machine into a training set, a validation set, and a test set based on a predetermined time and project interval. The training set is used to fit the Min–Max normalized parameters, as shown in formula (1). （1） in, The original value of the sample on the d-th feature. The normalized value of the sample on the d-th feature. Let be the minimum value of the d-th feature in the training set. Let be the maximum value of the d-th feature in the training set; Step A2 involves using the training set to perform normalization preprocessing on the multidimensional mechanism parameters, which include cutterhead speed, cutterhead torque, propulsion speed, penetration depth, total thrust, earth pressure, screw conveyor torque, and thrust factor. The normalization preprocessing includes label encoding, anomaly filling, and leakage prevention normalization.

3. The intelligent identification method for shield tunneling strata based on strict single-sample prototype network and validation set scale calibration as described in claim 2, characterized in that: Step B involves using a cosine-space classifier to perform supervised pre-training on the lightweight MLP encoder and obtain the discriminative embedding space. The specific steps are as follows: Step B1: Construct a lightweight MLP encoder and assign a cosine classification head, as shown in formula (2). ； ； （2） in, The L2-normalized sample embedding vector For MLP encoders, The input feature vector is the normalized form. As a learning standard, Cosine intervals, The class weight vector after L2 normalization. This is the original class weight vector; Step B2, calculate the weighted cross-entropy loss. As shown in formula (3), （3） in, For sample index, and All are category indexes. Let n be the true class label of the nth sample. and These are the numerator and denominator of the softmax function, respectively.

4. The intelligent identification method for tunnel boring machine strata based on strict single-sample prototype network and validation set scale calibration as described in claim 3, characterized in that: Step C involves selecting a unique supporting sample for each class from the training set within the embedding space based on a class-center-proximal sample strategy. The specific steps are as follows. Step C1: Embed and normalize the training set samples, then calculate the class mean for each class. As shown in formula (4), ； （4） in, Let be the normalized embedding vector of the nth training sample. This represents the selection of support samples for class c from the training set. For transpose; Step C2 involves selecting a unique supporting sample for each class, as shown in formula (5). （5） in, For class c, the fixed support prototype vector. Selected as supporting samples The normalized features of the training samples.

5. The intelligent identification method for tunnel boring machine strata based on strict single-sample prototype network and validation set scale calibration as described in claim 4, characterized in that: Step D involves keeping the support samples constant in each training task and sampling query samples from the training set in a class-balanced manner. An upper limit is set on the number of times a single sample is used in query samples within a single training period, and temperature-scaled cosine similarity cross-entropy is used to optimize the metric space. The specific steps are as follows. Step D1: In each training task, the support samples are kept unchanged and the query samples are sampled uniformly from the training set by class. During the sampling process, uniform sampling without replacement is used first, and padding with replacement is performed when there is insufficient sample. Step D2 sets an upper limit on the number of times a single sample can be used in a single training period, as shown in formula (6). （6） in, To query the number of times the sample was used, The maximum number of times each query sample is used within a single training task; Step D3 involves optimizing the metric space using temperature-scaled cosine similarity cross-entropy, as shown in Equation (7). ； （7） in, Let be the weighted cosine similarity of the i-th query sample with respect to category c. The first grid search scale, Let be the normalized embedding vector of the i-th query sample. The loss function for each training task, For query set, To query the sample index, For the i-th query sample in the true category The indexation results on, This represents the indexed result of the i-th query sample in the true class c.

6. The intelligent identification method for tunnel boring machine strata based on strict single-sample prototype network and validation set scale calibration as described in claim 5, characterized in that: Step E involves performing scale calibration on the validation set using the first grid search scale to obtain the second grid search scale, and then performing rigorous single-sample classification on the test set using the second grid search scale to obtain the classification results. The specific steps are as follows: Step E1: Perform scale calibration on the validation set using the first grid search scale to obtain the second grid search scale, as shown in formula (8). （8） in, This is the second grid search scale. For the candidate scale set, To improve classification accuracy, To verify the embedding vector, The category prototype matrix; Step E2: Perform strict single-sample classification on the test set using the second grid search scale and obtain the classification results, as shown in formula (9). （9） in, For the class label of the input sample x, Let x be the normalized embedding vector of sample x.

7. The intelligent identification method for tunnel boring machine strata based on strict single-sample prototype network and validation set scale calibration as described in claim 6, characterized in that: Step F involves processing the classification results using the softmax function to obtain the probability of the identified category, then averaging the probabilities using the identified category dimension and applying... The function obtains the final recognition result, as shown in formula (10). （10） in, For the final identification result, To identify the probability of the category.

8. A shield tunneling stratum intelligent identification system based on rigorous single-sample prototype network and validation set scale calibration, wherein the specific identification process of the shield tunneling stratum intelligent identification system is based on the shield tunneling stratum intelligent identification method according to any one of claims 1-7, characterized in that: It includes a preprocessing module, a cosine interval classification module, a fixed selection module, a balanced sampling module, a strict single-sample classification module, and a probability averaging module. The preprocessing module is used to divide the samples collected in real time from the tunnel boring machine into training set, validation set, and test set according to a predetermined time and engineering interval, and then use the training set to perform normalized preprocessing on the multidimensional mechanism parameters. The cosine interval classification module is used to perform supervised pre-training of the lightweight MLP encoder using a cosine interval classifier and obtain a discriminative embedding space. The fixed selection module is used to select a unique supporting sample for each class from the training set within the embedding space according to the class center proximal sample strategy. The equalization sampling module is used to keep the support samples unchanged in each training task and to equally sample query samples from the training set by class. It also sets an upper limit on the number of times a single sample is used in a single training period and uses temperature-scaled cosine similarity cross-entropy to optimize the metric space. The strict single-sample classification module is used to perform scale calibration on the validation set using a first grid search scale and obtain a second grid search scale, and then perform strict single-sample classification on the test set using the second grid search scale and obtain the classification result. The probability averaging module is used to process the classification results using the softmax function to obtain the probability of the identified category, and then to perform an arithmetic average using the identified category dimension and apply... The function obtains the final recognition result.

Images